Research
What the project's choices rest on: the question, the sources or measurements, what we decided, and the release it shipped in. Where something is not proven yet, it says so.
Measured means we ran it, and the entry says on what. Sourced means taken from published code or documentation. Open means we could not confirm it, it is a proposal, or it has not been run on real hardware.
Can the audio callback run without allocating, and is a 256-frame stream stable?
The render path is proven allocation-free under a test that aborts on any allocation, and a mutation check confirmed the guard is real. A 60 second stream with a 256-frame request ran with 0 xruns and 1 late callback.
The 60 second run was on a virtualised sink (WSLg), so the timing is not representative of real hardware, and the server may deliver in its own chunk size. In the spike, output latency was not measured. The second stream needed for headphone cue failed to open on that sink and is unproven on every backend. A run on native Linux hardware is still open.
- Keep cpal and own the render callback. Spike decisions
Which time-stretch engine keeps pitch and transients within budget on a stereo deck?
Several candidates were compared against a bar of under 5 cents of pitch error and under 5% of a core per stereo deck. No pure-Rust candidate passed. Rubber Band R3 held pitch within 3.4 cents but used 4.9 to 5.6% of a core, over the bar at one ratio. Signalsmith passed every measured bar and was the cheapest at 0.8 to 1.0%.
Measured on synthetic material only, as wall-clock on a shared WSL2 VM (plus or minus 20%). No real music, no stereo transient metrics and no listening check. Rubber Band R3 is over the CPU bar at one ratio, so the choice is to be re-measured in the real engine.
- Rubber Band R3 behind a small trait so it can be swapped. This was the author’s call, for reference quality, and may be revisited by ear. Its GPL-2.0-or-later licence is compatible with AGPLv3. Spike decisions
- In the codebase as om-stretch. Release not yet recorded.
Can a native GPUI window draw a scrolling custom waveform at 60 frames a second?
On WSL2 with X11, a window with a custom-painted scrolling waveform ran at 60.0 fps in a release build and 30.0 fps in debug, at about 114% CPU.
The window was never seen, and the figure is the app’s own frame counter. 60 and 30 look like a vsync or compositor cap, so this shows it keeps up with 60 but not how much headroom there is. The 30 fps debug reading is not understood. Software rendering is the likely cause of the CPU figure, and no real GPU was used. Wayland on WSLg failed at startup. These numbers are from gpui 0.2.2; the project has since moved to gpui-pre 0.3.10, so they describe the earlier version.
- Go on Linux; the gate is closed. Windows and macOS checks come with support for them. Spike decisions
- Release not yet recorded.
How should a waveform be coloured: by a hue wheel or by fixed frequency bands?
Mixxx colours its waveform from per-band levels blended on the GPU. It splits at 600 Hz and 4 kHz and stores the peak per stride rather than RMS, and Serato documents red for low, green for mid and blue for high. The research note recommends fixed band hues over a hue wheel.
No published crossover frequencies were found for Serato, Traktor, djay, rekordbox or Engine, so the 200 Hz and 3 kHz splits are our own choice and unvalidated against other programs. Light-theme blending is untested. Strip heights, playhead style and accessibility options could not be sourced.
- The note recommends three fixed band hues split at 200 Hz and 3 kHz, and leaves the final call open. Which style is the default is set in the later faceplate decisions. Faceplate decisions
- Release not yet recorded.
Which conventions should faders, knobs, deck controls and labels follow?
Interaction behaviour is sourced: for knobs, citing the FabFilter reference, vertical drag is primary, Shift gives fine control, and Ctrl or Cmd-click resets; eight hot cues per track is the norm. Everything here is conventions only, under the clean-room rule, with no artwork, layouts or text copied.
Pixel sizes, sweep angles and stroke widths are not published by any vendor, so they are our design choices to tune by eye. The eight hot cues claim is backed by one vendor guide, and a second source is a third-party summary not checked against the official manual. Some layout points come from working knowledge rather than a fetched page, and several figures are marked as proposals in the notes.
- Follow the sourced conventions where they fit and treat sizes and angles as our own design. Our knobs keep double-click as reset, with typed entry left for later. Knobs note
- Release not yet recorded.
Does the Rust stem pipeline match the reference Demucs model?
The changelog records a working htdemucs.onnx export, with the Rust pipeline checked against Demucs, in the first alpha.
Listening checks for stems are still open, and ONNX stem separation is behind an optional build feature. There is no separate research note for this yet, so the evidence is the pull request and the stems feature document.
- ONNX stem separation is an optional build feature, off by default, with its own CI job. Stems feature document